Unverified50% confidenceFactExact time
@Teknium found that Jev's compression tool score can be reproduced with one line of code (removing the tool call), and that the effect decays: the first round frees half the window, the next round frees a quarter, and then less.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/12/2026
First Seen
Valid until: 10/26/2026
Sources
Related Entities
Related Claims
UnverifiedIn the compression use case, a decision rule was set: when the session is below 25%, Jev is not activated and instead falls back to Claude's standard compression.63% similarUnverifiedThe jailbroken version removes the model's refusal layer by fine-tuning the original weights a second time or modifying the output layer bias, rendering the content filtering mechanism ineffective.63% similarUnverified整洁代码库与混乱代码库的通过率相近，但整洁代码消耗更少的token、导致更少的返工62% similarUnverifiedThe encoder-only repair approach can reduce trainable parameters by 3× and optimizer-state memory by 3×.61% similarUnverifiedWhen context usage reaches 70%–90%, it is best to compress the context or start a new conversation, because model performance degrades as it approaches the end of the context window.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1006163API
curl https://kongchang.com/api/v1/knowledge/claims/1006163MCP
get_claim(id=1006163)